Research And Application Of Dorsal Hand Vein Recognition Based On Multi-Feature Fusion

As a type of biometric recognition,dorsal hand vein recognition has received a lot of attention from researchers.Current research work focuses on traditional methods,such as extracting local binary pattern features from dorsal hand vein images and then inputting them into a specific classifier for recognition,which is a tedious process.With the prevalence of artificial intelligence,deep learning has demonstrated its powerful capabilities in the field of pattern recognition.However,the size of the dorsal hand vein dataset is small,and deep learning methods have not been intensively studied in this area.To address these problems,the main research works of this thesis are as follows:(1)For the problem of the small-scale of the dorsal hand vein dataset,we expand the dorsal hand vein dataset by adding Gaussian noise and using convolutional neural networks for dorsal hand vein recognition,which achieves a high recognition rate.Three convolutional neural networks and partition local binary pattern feature fusion methods are proposed for cross-database dorsal hand vein recognition.The method achieved good recognition results on the fused dorsal hand vein fusion dataset.(2)In order to improve the recognition results of small-scale samples of dorsal hand veins,we propose a method based on the feature fusion of the residual network and the histogram of gradients.The deep semantic features extracted from the residual network are fused with the image gradient information,and high recognition rates are achieved on three different small-scale sample datasets and the fused dataset.(3)In both of above methods,the dataset is expanded by Gaussian noise,but when conducting experiments,the recognition results are not satisfactory using the conventional method on the data set with Gaussian noise.This is because after adding Gaussian noise,the images of the same class appear to have a large difference in features,which is called intra-class variation.To address this problem,we propose a prior knowledge-guided adversarial generative network to expand the dorsal hand vein dataset.Meanwhile,in order to reduce the training time of the model,a large kernel convolutional residual network is proposed by fine-tuning the residual network,which reduces the training time by nearly half and compresses the model size to one-fourth of the original one while ensuring the accuracy of the model.Through the above study,the proposed method not only achieved high recognition rates on three different public datasets but also achieved state-of-the-art results on the fused dataset.